Over-Refusal and Representation Subspaces: A Mechanistic Analysis of Task-Conditioned Refusal in Aligned LLMs
For developers of aligned LLMs, this work provides a mechanistic understanding of over-refusal, highlighting the need for task-specific interventions rather than global ablation.
The paper analyzes why aligned LLMs over-refuse safe instructions that resemble harmful ones, finding that harmful-refusal directions are task-agnostic and captured by a single global vector, while over-refusal directions are task-dependent and span a higher-dimensional subspace. This explains why global refusal ablation fails to correct over-refusal and shows that task-specific geometric interventions are needed.
Aligned language models that are trained to refuse harmful requests also exhibit over-refusal: they decline safe instructions that seemingly resemble harmful instructions. A natural approach is to ablate the global refusal direction, steering the hidden-state vectors away or towards the harmful-refusal examples, but this corrects over-refusal only incidentally while disrupting the broader refusal mechanism. In this work, we analyse the representational geometry of both refusal types to understand why this happens. We show that harmful-refusal directions are task-agnostic and can be captured by a single global vector, whereas over-refusal directions are task-dependent: they reside within the benign task-representation clusters, vary across tasks, and span a higher-dimensional subspace. Linear probing confirms that the two refusal types are representationally distinct from the early transformer layers. These findings provide a mechanistic explanation of why global direction ablation alone cannot address over-refusal, and establish that task-specific geometric interventions are necessary.